Harnessing prototype networks for novel plant species and disease classification in open-world scenarios

Identifying novel plant categories and diseases in open-world scenarios is crucial for modern agricultural production and applications. Recent studies have extended plant disease recognition from closed-set to open-set scenarios, aiming to reject samples that do not belong to known classes. However, beyond rejection, it is necessary to classify unknown samples rather than merely labeling them as ”unknown.” This paper assumes that images of unknown samples, rejected by open-set recognition algorithms, are available. We aim to classify these unlabeled samples by leveraging prior knowledge from a labeled set. To the best of our knowledge, no existing research has addressed the classification of unknown plant species and diseases. To fill this gap, we propose a novel prototype network that models the category space relationship between known and unknown classes. Specifically, we learn a prototype vector for each known category, enabling samples to obtain distance-based category probabilities by measuring their similarity to these prototypes. This approach captures complex class boundaries more effectively than linear classification models, offering greater flexibility and accuracy. Additionally, we employ a knowledge distillation loss to optimize the category space relationship and calculate a consistency loss to balance the model's classification performance for both known and unknown classes. To further boost performance, we incorporate the pre-trained model dino-v2. Experiments on the large-scale plant specimen dataset Herbarium19 and the plant disease dataset Plant Village demonstrate that our method surpasses baseline approaches, improving novel class accuracy by 1%–30%. This research contributes to intelligent agriculture, and we will release the code to facilitate future work in the field.